Thanks for sharing these @Liviacmg. I think for those samples, in which you're losing more sequences, appears to be due to losing proportionally more due quality filtering rather than chimera removal. The drop from denoised to non-chimeric isn't as much of a drop as the quality filtering. This could probably be helped by truncating a little more.
Another option is to adjust the --p-min-fold-parent-over-abundance to 8 or 16. I wouldn't suggest any higher than that. This is based on the following:
Some of these refer to usearch / vsearch parameters, which are analogous to this DADA2 parameter. Much of this is explored in the github thread.